Method for training multi-class classifiers with active selection and binary feedback

1. A method for training a multi-class classifier, comprising the steps of:

selecting a query image from a set of active images based on a membership probability determined by the classifier, wherein the active images are unlabeled;

selecting a sample image from a set of training image based on the membership probability of the query image, wherein the training images are labeled;

displaying the query image and the sample images to a user on an output device;

obtaining a response from the user with an input device, wherein the response is a yes-match or a no-match;

adding the query image with the label of the sample image to the training set if the yes-match is obtained, and otherwise repeating the selecting, displaying, and obtaining steps until a predetermined number of no-match is reached to obtain the multi-class classifier;

applying the multi-class classifier to a set of unlabeled images to obtain a set of detection results;

determining membership probabilities of the set of detection results;

associating the set of detecting results with the membership probabilities less than a predetermined threshold with the set of active images; and

retraining the multi-class classifier to refine the set of detection results, wherein steps of the method are performed by a processor.

2. The method of claim 1 , further comprising:

estimating the membership probability using a one-versus-one support vector machines, and sequential minimal optimization.

3. The method of claim 2 , wherein the membership probability is estimated using logistic regression classifiers.

4. The method of claim 2 , wherein the membership probability is estimated using Fisher discriminant analysis over multiple classes.

5. The method of claim 1 , further comprising:

assigning a new class label to the query image if the predefined number of no-match is reached.

6. The method of claim 1 , wherein the initial set of training image are selected randomly from a seed set of images.

7. The method of claim 1 , wherein the initial set of training image is selected by querying a database of an image search engine.

8. The method of claim 1 , wherein the initial set of training image is supplied by the user.

9. The method of claim 1 , wherein a plurality of users are used to obtain a plurality of responses.

10. The method of claim 1 , wherein the selecting of the query image is based on an expected value of information (EVI), with an objective function that combines an expected risk, and a cost of user labeling, and the selecting of the sample image is based on maximum likelihood of a match with the query image.

11. The method of claim 1 , further comprising:

merging of trained classes by an agglomerative clustering.

12. The method of claim 11 , further comprising:

combining two classes if a between-class similarity score is larger than a predetermined threshold within the agglomerative clustering.

13. The method of claim 11 , further comprising:

combining two clusters if the response is the yes-match when the two classes are displayed to the user, and asking whether the two classes should be merged or not.

14. The method of claim 1 , further comprising:

adjusting the expected risk factors depending on an application and domain specifications.

15. The method of claim 1 , further comprising:

obtaining a disregard response from the user to delete the query image from set of training images.